One malformed detection field aborts the entire scan's insert

**File:** src/db/postgres.py:89 (feature_to_row / insert_detections)

`feature_to_row()` does unguarded `float()`/dict-key access for every detection inside `insert_detections()`'s list comprehension, which is built before any DB call.

**Failure scenario:** A single detection in a fetch has an empty/non-numeric `bright_ti4` or `frp` field, or is missing a key. The list comprehension raises `ValueError`/`KeyError` before `executemany` ever runs. This propagates uncaught out of `reload_data()`, so `cache.set_current()` never runs and every other valid detection in that same fetch is dropped along with it.

**Category:** correctness

Start in src/db/postgres.py at feature_to_row and insert_detections, then trace how reload_data handles their result. Verify the behavior with a fetch containing one malformed or missing detection field: valid detections should still be inserted, and reload_data should continue far enough for cache.set_current() to run.

Written by the indexing model from the issue text.